Machine learning approaches to predict early cardiac immune-related adverse events in patients receiving immune checkpoint inhibitors.

PURPOSE: Immune checkpoint inhibitor (ICI)-induced cardiac immune-related adverse events (cardiac irAEs) are rare yet serious complications. Clinical assessment tools to identify at-risk patients would allow for more effective prevention strategies, thus improving clinical outcomes. We constructed various machine learning (ML) models to predict these events among patients receiving ICI therapy. METHODS: A cohort of patients receiving ICI therapy from 2010 to 2023 was identified from the TriNetX database. Cardiac irAEs were defined as the occurrence of relevant diagnosis codes within 90 days of ICI initiation, with corresponding hospital visits. We created ML models to predict these events, including elastic net logistic regression and multiple tree-based approaches (gradient boosted trees and random forest). We evaluated model performance with different performance measures and utilized assigned risk scores to stratify risk of cardiac irAEs into low, medium, and high-risk tiers. RESULTS: We identified 61,117 patients receiving ICI therapy, with nearly 2% of patients experiencing cardiac irAEs. Model performance on testing data was comparable with all approaches (AUC = 0.71-0.72, balanced accuracy = 65-66%). Each model emphasized distinct features to make classifications, as observed with feature importance and SHAP values. Comparing cardiac irAE rates among assigned risk strata, patients identified as high risk were significantly more likely to experience cardiac irAEs compared to lower tiers. CONCLUSION: Our preliminary exploration of ML methods demonstrated the potential for risk assessment tools to predict rare cardiac irAEs in patients receiving ICI therapy. Follow-up studies can implement time series approaches to harness longitudinal data that incorporates real-time labs, new diagnoses, and new therapy, to refine predictions further.